Estimating Subsurface Chlorophyll‐a Vertical Structure in Tropical‐Subtropical Oceans Using TCB‐MHA: A Dual‐Branch Neural Network Model

Abstract The vertical distribution of subsurface chlorophyll (Chla) is crucial for understanding the marine primary productivity estimation and the ocean carbon cycle. Here we developed a dual‐branch neural network (TCB‐MHA) that integrates satellites, reanalysis, and BGC‐Argo data to accurately estimate Chla profiles in tropical‐subtropical oceans. The model shows exceptional performance in estimating subsurface Chla concentration ( R 2 = 0.824–0.859, RMSE = 0.0525–0.0980 mg/m 3 ), capturing the depth ( R 2 = 0.919, MAPE = 7.23%) and intensity ( R 2 = 0.795, MAPE = 12.97%) of the subsurface chlorophyll maximum (SCM). However, SCM thickness prediction remains challenging ( R 2 = 0.513, MAPE = 16.82%), likely due to the lack of phytoplankton photoacclimation parameterization. Independent validation in the South China Sea confirmed the model's strong generalizability. Climatological analysis reveals clear latitudinal SCM patterns across tropical‐subtropical oceans: deeper and weaker in subtropical regions (SCM depth >100 m, SCM intensity = 0.32 ± 0.10 mg/m 3 ) versus shallower and stronger in tropical zones (SCM depth <80 m, SCM intensity = 0.48 ± 0.10 mg/m 3 ). Seasonal and vertical SCM variations are also consistently captured. The model further quantifies the impact of mesoscale processes and spring blooms of the Western Mediterranean on Chla distribution. This study highlights the advantages of deep learning in integrating multi‐source heterogeneous ocean data, and the resulting monthly 3D Chla products can be applied to improve carbon cycle and productivity studies.